Cultural evolution in populations of Large Language Models

TL;DR

Problem: understanding how cultural traits emerge, spread, and stabilize over time. Approach: simulate linguistic cultural evolution using networks of generative AI agents (LLMs) guided by initialization, transformation prompts, and personality settings, with network topology and social aggregation manipulated. Contributions: an open-source software toolkit with TF-IDF cosine similarity analysis, word-chain and similarity-network visualizations, and systematic variation of network structure, prompts, and personalities. Significance: demonstrates that LLM-based multi-agent models can reproduce known cultural-dynamics patterns and provide a versatile platform to study both human and machine-generated culture, bridging cultural evolution and generative AI research.

Abstract

Research in cultural evolution aims at providing causal explanations for the change of culture over time. Over the past decades, this field has generated an important body of knowledge, using experimental, historical, and computational methods. While computational models have been very successful at generating testable hypotheses about the effects of several factors, such as population structure or transmission biases, some phenomena have so far been more complex to capture using agent-based and formal models. This is in particular the case for the effect of the transformations of social information induced by evolved cognitive mechanisms. We here propose that leveraging the capacity of Large Language Models (LLMs) to mimic human behavior may be fruitful to address this gap. On top of being an useful approximation of human cultural dynamics, multi-agents models featuring generative agents are also important to study for their own sake. Indeed, as artificial agents are bound to participate more and more to the evolution of culture, it is crucial to better understand the dynamics of machine-generated cultural evolution. We here present a framework for simulating cultural evolution in populations of LLMs, allowing the manipulation of variables known to be important in cultural evolution, such as network structure, personality, and the way social information is aggregated and transformed. The software we developed for conducting these simulations is open-source and features an intuitive user-interface, which we hope will help to build bridges between the fields of cultural evolution and generative artificial intelligence.

Figures (20)

• Figure 1: (a) LLM agents are organized into networks wherein each agent interacts with neighboring agents by exchanging stories. (b) Each agent is assigned a specific personality and either initialization instructions (for the first generation) or transformation instructions (after the first generation), serving as prompts for generating new stories from their neighbors' narratives. (c) Once the network structure and agent characteristics are defined, we simulate the cultural evolution of texts across generations of agents. The simulation begins by prompting agents to initialize stories, after which we allow the narratives to evolve dynamically through interactions within the agent network
• Figure 1: Different types of networks used in our experiments with 10 agents. In a fully connected network (a), every agent is directly connected to every other agent, enabling efficient dissemination of information. In a circular network (b), agents are connected in a circular fashion, thereby forming a closed loop where stories flow sequentially around the circle. The caveman network (c) consists of agents organized into cliques, with our example showcasing two cliques of 5 agents each. In the chain network (d), agents are arranged linearly, and content generation happens sequentially, meaning that agents only generate a story after receiving the story of the previous agent in the chain.
• Figure 2: Visualization of the evolution of the texts generated along a chain of 50 agents (i.e. 50 generations of one agent per generation). (a) The similarity matrix represents the semantic similarity between all stories generated. The color of the cell at row i, column j corresponds to the similarity of the stories i and j, which here corresponds to the stories generated at generation i and generation j. (b) The similarity graph is another way of visualizing the similarity between all generated stories. Each node corresponds to one story, and the distance between node is proportional to the semantic distance between corresponding stories. Stories generated at successive generations are linked by a wider edge and arer assigned similar colors. (c) We also visualize the words used at each generation. Words are along the x-axis, and generation along the y-axis. A dot at position (x,y) means that the word x was used in a story at generation y. A line means that the corresponding word was used by two successive generations. We only display the first half of the figure here. The complete figure can be found in the Fig. \ref{['rotated_word_chains']} of Appendix.
• Figure 2: Screenshot of the Graphical User Interface
• Figure 3: Effect of network structure : Cultural dynamics of a population of 10 agents over 10 generations for three different types of network structures. (a), (b), (c) Similarity matrices for a fully-connected network, a caveman network with 2 cliques, and a circle network. The index of a story is defined as the agent's index (here between 0 and 10) + $N_{agents}$ * generation-index. For example, the story 3 belongs to the first generation, and story 13 to the second generation. The color of the cell at row i and column j represents the semantic similarity between stories i and j. Black lines mark the separation between generations. (d) Evolution of the average similarity between stories produced at a generation and stories at the generation just before.(e) Evolution of the average similarity between each pair of stories produced at a given generation.(f) Evolution of the average similarity between stories produced a given generation and stories produced at the first generation. Lines represent averages over 5 simulations, and the filled areas represent the standard deviations.